Conventionally, well known is a technique using sound source separation of blind sound source separation method according to an independent component analysis method in a frequency region on sound source signals outputted from a plurality of sound sources, to generate respective separation signals corresponding to the sound source signals from a plurality of mixed sound source signals obtained by superimposing the sound source signals (e.g., Patent Documents 1 to 3).
In a technique of the Patent Document 1, by the sound source separation of blind sound source separation method according to the independent component analysis method in a frequency region, a SIMO (Single-Input Multiple-Output) signal is generated as a plurality of separation signals for each frequency bin. Next, among the plurality of separation signals, a first separation signal corresponding to a sound source to be separated and second separation signals other than the separation signal corresponding to this sound source are compared for each frequency bin. Then, by mask processing on the basis of a result of comparison among these separation signals, a noise component is removed from the first separation signal for each frequency bin and a target signal is thereby generated.
In a technique of the Patent Document 2, by utilizing the fact that an arrival direction of the sound source signal outputted from the sound source to be separated and an arrival direction of the noise signal are different from each other, the sound source separation is performed. Specifically, after the sound source separation according to the independent component analysis method in a frequency region, a cross correlation between a separation signal of straight component corresponding to the target signal and a separation signal of cross component corresponding to an interfering sound is calculated and a coefficient for noise estimation is obtained from the amount of delay of when the cross correlation becomes the maximum. Then, on the basis of this obtained coefficient, a noise component is removed from the separation signal corresponding to the target signal.
Further, in a technique of the Patent Document 3, noise estimation and noise removal are performed on the basis of the assumption that an amplitude spectrum of the sound source signal outputted from the target sound source and that of the noise signal do not simultaneously take large values at the same time in the same frequency.    Patent Document 1: Japanese Patent Application Laid Open Gazette No. 2006-154314    Patent Document 2: Japanese Patent Gazette No. 3831220    Patent Document 3: Japanese Patent Application Laid Open Gazette No. 2005-308771